Bidding-Aware Retrieval (BAR) in Online Advertising

• BAR is a design principle that incorporates bid context into retrieval scoring to reduce multi-stage inconsistencies in ad ranking.
• It employs techniques like Bidding-Aware Objectives and Task-Attentive Refinement to optimize performance under dynamic bidding environments.
• BAR frameworks extend to applications in generative recommendation and auto-bidding, balancing revenue maximization, relevance, and risk management.

Bidding-Aware Retrieval (BAR) denotes a class of retrieval and candidate-selection mechanisms in which bids, bidding context, or an explicit downstream utility proxy are incorporated into retrieval itself rather than deferred to a later ranking or auction stage. In industrial display advertising, BAR is a model-based retrieval framework that incorporates ad bid value into the retrieval scoring function to reduce the inconsistency between bid-agnostic retrieval and eCPM-based downstream ranking (Liu et al., 7 Aug 2025). Closely related formulations appear in offline auto-bidding, where retrieval supplies high-quality historical bid candidates for value-based selection; in generative recommendation, where bids are injected directly into semantic-ID decoding; in content promotion, where bids encode both short-term value and information gain; and, in a utility-aware interpretation, in retrieval-augmented generation, where evidence selection is optimized for generator utility rather than pure relevance (Cui et al., 12 Jun 2026, Jiang et al., 23 Mar 2026, Liu et al., 28 Jan 2026, Sun et al., 3 Feb 2026).

1. Scope, definitions, and recurrent problem structure

Across the cited literature, BAR addresses a recurring mismatch between a lightweight retrieval stage and a later decision stage with a richer objective. In display advertising, the mismatch is explicitly described as multi-stage inconsistency: retrieval cannot access precise, real-time bids for the vast ad corpus, while pre-ranking, ranking, and re-ranking allocate traffic according to eCPM or closely related objectives (Liu et al., 7 Aug 2025). In offline auto-bidding, the analogous failure mode is the Average Action trap, together with unreliable behavior under sparse/long-tail traffic, when a purely parametric policy compresses multiple valid bidding modes into a single averaged action (Cui et al., 12 Jun 2026). In generative recommendation, the limitation is that preference-only semantic-ID models do not distinguish organic from sponsored modes and cannot react to real-time bids without retraining (Jiang et al., 23 Mar 2026). In RAG, the corresponding critique is that retrievers and rerankers optimize solely for relevance, not for whether evidence is suitable for the generator (Sun et al., 3 Feb 2026).

Setting	BAR mechanism	Representative work
Display advertising retrieval	eCPM-oriented retrieval score with bid features, monotonicity constraints, and near-line embedding updates	BAR (Liu et al., 7 Aug 2025)
Offline auto-bidding	retrieval-augmented candidate generation from historical decisions, followed by critic-based selection	DRIVE (Cui et al., 12 Jun 2026)
Generative recommendation	bid-aware decoding over semantic IDs with control tokens for sponsored vs. organic mode	GEM-Rec (Jiang et al., 23 Mar 2026)
Retrieval-augmented generation	boundary-aware evidence selection trained from generator feedback	BAR-RAG (Sun et al., 3 Feb 2026)
Content promotion	bids derived from short-term value and long-term information gain	(Liu et al., 28 Jan 2026)
Risk-aware RTB	risk-adjusted bid scores based on CTR and auction uncertainty	(Zhang et al., 2017)

This suggests that BAR is best understood not as a single architecture but as a design principle: retrieval is made aware of the quantity that ultimately governs utility, whether that quantity is eCPM, return-to-go, revenue, gradient coverage, or generator reward.

2. BAR in multi-stage online advertising systems

The most literal use of the term appears in "Bidding-Aware Retrieval for Multi-Stage Consistency in Online Advertising" (Liu et al., 7 Aug 2025). The paper studies a cascaded architecture with Retrieval, Pre-Ranking, Ranking, and Re-Ranking, and formulates the platform objective as selecting a subset of ads to maximize expected revenue under per-ad

$$\mathrm{eCPM}_i = \mathrm{pCTR}_i \times \mathrm{pBid}_i \times 1000.$$

The central claim is that retrieval is historically bid-agnostic or only loosely bid-aware because it must operate over tens of millions of ads under strict latency and memory limits, typically with pre-computed ANN embeddings that are updated much more slowly than auto-bidding signals.

BAR replaces CTR-oriented or similarity-based retrieval with a learned scalar score

$f_{\mathrm{eCPM}}(u,a),$

used in a point-wise top-$k$ retrieval rule. The model is trained with a pairwise learning-to-rank objective over ranking-stage exposure logs, so that ads shown by downstream ranking are preferred to non-exposed or randomly sampled alternatives. The framework then adds Bidding-Aware Modeling, composed of a Bidding-Aware Objective (BAO) and a Distillation Auxiliary Objective (DAO). BAO enforces correct monotone responses to changes in bid-related features such as budget_left_ratio and bid_constraint; DAO distills downstream pCTR and pBid into auxiliary heads. The total objective is

$$\mathcal{L}_{\text{total}} = \mathcal{L}_{\text{LTR}} + \lambda_1 \mathcal{L}_{\text{BAO}} + \lambda_2 \mathcal{L}_{\text{DAO}}.$$

A second defining component is Asynchronous Near-Line Inference. The deployed model is split into a User-Side Graph and an Ad-Side Graph. Full-batch ad embeddings are indexed with HNSW for online retrieval, while a near-line service listens to bid-constraint and budget-change events, recomputes affected ad embeddings, and asynchronously updates index entries. The paper emphasizes that this yields embedding updates within seconds of bid changes, with millions of read QPS and thousands of write QPS, using fine-grained readers-writer spinlocks at per-ad entry granularity.

BAR also introduces Task-Attentive Refinement (TAR) to disentangle user interest from commercial value. The pCTR head cross-attends to user behavior sequence features; the pBid head cross-attends to raw ad features, especially bid-related ones; and the eCPM head fuses hidden representations from both branches with a projection of the joint user-ad representation. This is meant to avoid the large user embedding overwhelming dynamic bidding signals.

The reported empirical evidence is both offline and online. Offline, on Alibaba display ads, full BAR on MBR achieves Recall@2000 = 54.9% and PairAcc = 96.1%, compared with 52.6% and 80.2% without BAO, and 50.9% and 50.0% without bid features. TAR raises Recall-all@2000 to 68.2% with 1.2× GFLOPs, compared with 64.3% for the baseline and 66.8% for the +IncDim variant at 4× GFLOPs. In a one-month A/B test on Alibaba display advertising, BAR yields +4.32% platform revenue, +3.78% RPM, +0.31% CTR, and +0.01% ROI, while responsiveness to positive advertiser operations rises sharply: in the “Ad Value” channel, RIR increases from 2.5% to 30.4%, and across all channels IIR increases by 22.2% (Liu et al., 7 Aug 2025).

3. Action-level BAR in offline auto-bidding

"DRIVE: Distributional and Retrieval-Augmented Bidding with Value Evaluation" presents BAR as the core of an offline auto-bidding framework for real-time advertising (Cui et al., 12 Jun 2026). The environment is a generalized second-price (GSP) auction over impression opportunities with values $v_i$, payments $c_i$, and win indicators $x_i$, under a budget $B$ and optional KPI constraints. Under mild assumptions, optimal impression-level bids have the scaled form

$b_i^* = \lambda v_i,$

so the control problem is reduced to choosing a scalar bidding parameter $\lambda_t$ over a campaign day modeled as an MDP.

DRIVE identifies two pathologies in Transformer-style offline RL or sequence modeling. The first is the Average Action trap: similar states may support multiple effective bidding modes, such as aggressive and conservative pacing, but unimodal policies regress toward a single suboptimal mean. The second is failure under Sparse/long-tail traffic, where purely parametric policies hallucinate unreliable actions in low-density regions even though high-quality historical decisions exist in the log. BAR is introduced precisely to address these failure modes.

The retrieval mechanism is built over an offline trajectory dataset. Each time step is encoded into a contextual state embedding

$f_{\mathrm{eCPM}}(u,a),$0

and the index stores the action $f_{\mathrm{eCPM}}(u,a),$1 and return-to-go $f_{\mathrm{eCPM}}(u,a),$2. At inference time, the current context is encoded, approximate nearest neighbors are retrieved by cosine similarity using a FAISS HNSW index, a larger similarity-based pool is formed, and then a bidding-aware filter keeps the $f_{\mathrm{eCPM}}(u,a),$3 entries with highest stored RTG. Retrieved actions are therefore both state-similar and high-quality. The paper states that embeddings are normalized and that search uses inner product ≈ cosine similarity.

DRIVE does not use retrieval alone. It decouples candidate generation from decision making. A Transformer with a Gaussian Mixture Model (GMM) head models a multi-modal conditional action distribution and samples generated candidates; BAR contributes retrieved historical actions; and an offline Implicit Q-Learning (IQL) critic evaluates all candidates and selects

$f_{\mathrm{eCPM}}(u,a),$4

The training procedure remains modular: the policy is trained by GMM maximum likelihood on trajectories, the critic by $f_{\mathrm{eCPM}}(u,a),$5 and $f_{\mathrm{eCPM}}(u,a),$6, and the retrieval index is constructed offline after encoder training.

The paper reports that approximately 17.6% of states have multimodal $f_{\mathrm{eCPM}}(u,a),$7, with a suboptimality rate of approximately 55% on multimodal states versus approximately 38% on unimodal states. In ablations, Actor only is worst, GMM + critic improves substantially, Retrieval + critic helps but is limited, and full DRIVE is best across budgets. On AuctionNet, at 150% budget, DRIVE obtains 551 ± 4.64 versus 535 ± 2.97 for the best baseline. On AuctionNet-Sparse, at 100% budget, DRIVE reaches 37.3 ± 0.87 versus 30.6 ± 0.69 for DT. The paper also emphasizes deployability: GMM + critic runs at approximately 10–11 ms per step versus approximately 223 ms for diffusion, and full DRIVE runs at approximately 46 ms per decision versus approximately 10 ms for DT, still within typical RTB latency budgets (<50 ms) (Cui et al., 12 Jun 2026).

4. Bid-aware generative retrieval in recommender systems

"One Model, Two Markets: Bid-Aware Generative Recommendation" adapts BAR to semantic-ID generative recommendation (Jiang et al., 23 Mar 2026). Each item is represented by a hierarchical semantic code

$f_{\mathrm{eCPM}}(u,a),$8

learned by an RQ-VAE. GEM-Rec augments this representation with control tokens

$f_{\mathrm{eCPM}}(u,a),$9

so that the model first decides whether the next slot is organic or sponsored and then generates the item’s semantic ID conditioned on that mode. The generative factorization is

$k$0

The training logs contain only successful placements, which the paper interprets as a learned feasibility policy: contexts where ads were historically both monetizable and acceptable to users.

The BAR mechanism is Bid-Aware Decoding, applied entirely at inference time. At the slot level, the sponsored-flag logit is shifted by the maximum bid among eligible ad candidates,

$k$1

while the organic-flag logit is unchanged. At the item level, semantic-ID decoding uses a prefix-level bid aggregator

$k$2

and token logits are modulated as

$k$3

Beam search is then run conditionally: if the slot is organic, decoding uses unmodulated logits; if it is sponsored, decoding uses the modulated logits.

A central contribution is the economic characterization. Proposition 1 (Allocative Monotonicity) states that, for fixed competing bids, the exposure probability of ad $k$4 is non-decreasing in its bid $k$5. Proposition 2 (Structural Consistency) states three properties: Safe fallback when $k$6, Organic integrity, and Generalization to all-organic data. Organic integrity is especially important: increasing $k$7 can change how often organic slots occur, but not the relative ranking among organic items.

The experiments use four public datasets — Steam, Amazon Beauty, Amazon Sports, and Amazon Toys — with a synthetic marketplace. On the Steam example, with train ad percentage approximately 3.5%, TIGER has Ad rate 0% and NDCG@10 ≈ 0.1442; GEM-Rec $k$8 has Ad rate 2.5%, Revenue 535, NDCG@10 ≈ 0.1411, and Organic NDCG@10 ≈ 0.1468; GEM-Rec $k$9 has Ad rate 4.7%, Revenue 1173, NDCG@10 ≈ 0.1381, and Organic NDCG@10 ≈ 0.1467. The stated pattern is that Revenue rises strongly with $$\mathcal{L}_{\text{total}} = \mathcal{L}_{\text{LTR}} + \lambda_1 \mathcal{L}_{\text{BAO}} + \lambda_2 \mathcal{L}_{\text{DAO}}.$$0, Ad rate increases smoothly, Total NDCG declines as more ads are inserted, but Conditional Organic NDCG is nearly flat. Under bid shocks that raise bids by 10× for a random 5% subset of items, the model rapidly shifts toward high-bid items; on Steam at $$\mathcal{L}_{\text{total}} = \mathcal{L}_{\text{LTR}} + \lambda_1 \mathcal{L}_{\text{BAO}} + \lambda_2 \mathcal{L}_{\text{DAO}}.$$1, High-Value Share jumps from 21.8% to 81.5% with a small increase in ad rate and revenue ≈ $$\mathcal{L}_{\text{total}} = \mathcal{L}_{\text{LTR}} + \lambda_1 \mathcal{L}_{\text{BAO}} + \lambda_2 \mathcal{L}_{\text{DAO}}.$$2 (Jiang et al., 23 Mar 2026).

5. Utility-aware, information-aware, and risk-aware generalizations

A broader line of work uses BAR-like principles to optimize retrieval or bidding against downstream utility rather than a static relevance score. "Guiding the Recommender: Information-Aware Auto-Bidding for Content Promotion" formulates content promotion as a dual-objective problem

$$\mathcal{L}_{\text{total}} = \mathcal{L}_{\text{LTR}} + \lambda_1 \mathcal{L}_{\text{BAO}} + \lambda_2 \mathcal{L}_{\text{DAO}}.$$3

where $$\mathcal{L}_{\text{total}} = \mathcal{L}_{\text{LTR}} + \lambda_1 \mathcal{L}_{\text{BAO}} + \lambda_2 \mathcal{L}_{\text{DAO}}.$$4 is expected short-term value and $$\mathcal{L}_{\text{total}} = \mathcal{L}_{\text{LTR}} + \lambda_1 \mathcal{L}_{\text{BAO}} + \lambda_2 \mathcal{L}_{\text{DAO}}.$$5 is gradient coverage, a surrogate for long-term model improvement with a formal connection to Fisher Information and optimal experimental design (Liu et al., 28 Jan 2026). The paper proves that the composite objective is monotone submodular, derives a two-stage Lagrangian auto-bidding algorithm, and gives the impression-level bid rule

$$\mathcal{L}_{\text{total}} = \mathcal{L}_{\text{LTR}} + \lambda_1 \mathcal{L}_{\text{BAO}} + \lambda_2 \mathcal{L}_{\text{DAO}}.$$6

with the second-price simplification

$$\mathcal{L}_{\text{total}} = \mathcal{L}_{\text{LTR}} + \lambda_1 \mathcal{L}_{\text{BAO}} + \lambda_2 \mathcal{L}_{\text{DAO}}.$$7

To handle missing labels at bid time, it introduces a confidence-gated gradient heuristic and a zeroth-order (ZO) variant for black-box models. Empirically, the proposed method achieves the highest AUC and lowest LogLoss in end-to-end offline promotion experiments, while budget pacing tracks targets closely (Liu et al., 28 Jan 2026).

BAR-RAG extends the same intuition to evidence selection in retrieval-augmented generation (Sun et al., 3 Feb 2026). The paper explicitly reframes reranking as a boundary-aware evidence selector targeting the generator’s Goldilocks Zone, defined by evidence sets with empirical correctness probability

$$\mathcal{L}_{\text{total}} = \mathcal{L}_{\text{LTR}} + \lambda_1 \mathcal{L}_{\text{BAO}} + \lambda_2 \mathcal{L}_{\text{DAO}}.$$8

The selector is trained with reinforcement learning, using a reward that combines a boundary term, a relevance term, a format reward, and a count penalty, while the generator is then fine-tuned under the selector-induced evidence distribution to mitigate train-test mismatch. The paper reports an average gain of 10.3 percent over strong RAG and reranking baselines and, for Qwen2.5-7B, an average 39.1 EM versus 26.9 for RAG. Although BAR-RAG is not about auction bids, the paper’s own interpretation is that it realizes a utility-aware retrieval mechanism in which evidence is selected for downstream utility rather than raw relevance. This suggests a generalized BAR perspective in which the “bid” is the expected contribution of a candidate set to generator performance.

A complementary precursor relevant to BAR design is "Managing Risk of Bidding in Display Advertising" (Zhang et al., 2017). That paper models uncertainty in both CTR and market price, defines a VaR-style risk-adjusted utility

$$\mathcal{L}_{\text{total}} = \mathcal{L}_{\text{LTR}} + \lambda_1 \mathcal{L}_{\text{BAO}} + \lambda_2 \mathcal{L}_{\text{DAO}}.$$9

and derives two risk-aware bidding strategies: VaR-based bidding and Risk Management of Profit (RMP). The resulting BAR-relevant score

$v_i$0

penalizes impressions with high CTR uncertainty. The empirical study reports profit gains of 15.4% in offline experiments and up to 17.5% in an online A/B test. In a BAR interpretation, these strategies supply risk-adjusted retrieval or ranking signals that internalize uncertainty in user response and auction competition (Zhang et al., 2017).

6. Limitations, misconceptions, and research directions

The literature presents several recurring misconceptions. First, BAR is not merely a post-hoc reranker. In industrial advertising, it changes the retrieval score itself and maintains a dynamic ANN index (Liu et al., 7 Aug 2025). In DRIVE, retrieval contributes directly to action candidate generation and is combined with critic-based selection at inference time (Cui et al., 12 Jun 2026). In GEM-Rec, bid information is injected into the generative decoding process rather than added after candidate generation, and organic integrity ensures that organic ordering is unaffected by bid modulation within organic slots (Jiang et al., 23 Mar 2026). Second, BAR is not equivalent to always favoring the highest bidder. The cited systems consistently fuse bid signals with user-interest or task-utility signals: BAR’s eCPM head distills pCTR and pBid jointly, DRIVE filters by RTG and evaluates with a conservative critic, GEM-Rec conditions ad generation on semantic plausibility, and information-aware promotion explicitly balances short-term value with long-term learning value (Liu et al., 7 Aug 2025, Cui et al., 12 Jun 2026, Jiang et al., 23 Mar 2026, Liu et al., 28 Jan 2026).

The main limitations are also recurrent. DRIVE assumes offline dataset coverage and approximate stationarity, and reports higher memory and latency than a plain DT baseline: approximately 29 GB RAM, with approximately 13 GB due to the FAISS index, and approximately 46 ms per decision (Cui et al., 12 Jun 2026). Industrial BAR requires a complex near-line serving stack, dynamic index mutation, and significant offline training cost; the deployed MBR variant meets latency at approximately 500 QPS at 40 ms, versus 600 QPS for the baseline (Liu et al., 7 Aug 2025). GEM-Rec does not guarantee DSIC, uses first-price auctions in experiments, and is effectively single-slot with a synthetic marketplace (Jiang et al., 23 Mar 2026). BAR-RAG is computationally expensive because selector training requires generator rollouts, uses 8×A100, and depends on a generator strong enough for competence-boundary learning (Sun et al., 3 Feb 2026). Information-aware promotion relies on gradient proxies or ZO estimators and is validated offline on synthetic and Criteo data rather than in a live exchange (Liu et al., 28 Jan 2026).

The cited future directions are correspondingly diverse. Industrial BAR suggests extension to other ad formats, GMV-weighted or ROI-constrained objectives, and more sophisticated monotonic architectures (Liu et al., 7 Aug 2025). GEM-Rec identifies DSIC-compatible or approximately truthful payment rules, multi-slot, multi-objective settings, and application to real-world ad logs as open problems (Jiang et al., 23 Mar 2026). DRIVE suggests that BAR-style retrieval plus distributional policy plus value evaluation is not tied to a specific Transformer design, since augmenting BC, CDT, PDiT, and DT all yields gains (Cui et al., 12 Jun 2026). BAR-RAG points toward retrieval systems that optimize expected downstream reward under explicit constraints rather than relevance alone (Sun et al., 3 Feb 2026). Taken together, these directions indicate that BAR is evolving from a narrowly advertising-specific retrieval adjustment into a more general principle for coupling candidate selection to the quantity that actually determines utility at decision time.